How our exit-intent ML model actually works: a deep dive for marketers in 2026

Beyond the Mouse: The 26 Signals Our Popup ML Watches

When we talk about exit-intent, most people picture a mouse cursor moving out of the browser window. While that's one signal, it's just the tip of the iceberg. Our ExitSense ML model continuously monitors 26 distinct behavioral signals to predict a user's intent to leave. These signals encompass everything from scroll velocity and direction to idle time, tab switching, form interaction, and even subtle micro-movements of the pointer.

For example, a user who rapidly scrolls to the bottom of a page, then quickly scrolls back up, pauses, and then hovers near a prominent navigation element might be signaling a specific kind of 'lost intent.' Each signal contributes a weighted value to the model, creating a dynamic user profile that updates in real-time. This allows us to move beyond simple rule-based triggers and instead respond to nuanced user journeys. This granular understanding is key to how our exit-intent ML model actually works.

Thompson Sampling Explained for Marketers: Why We Don't Just A/B Test

Traditional A/B testing is valuable but can be slow, especially for optimizing elements like popup headlines where many variations are possible. That's why we employ Thompson sampling. Instead of splitting traffic 50/50 and waiting for statistical significance, Thompson sampling dynamically allocates more traffic to variations that are performing better, faster. It's a 'multi-armed bandit' approach where the system learns and adapts in real-time.

Think of it like a casino player trying different slot machines: Thompson sampling identifies the 'winning' machines (headlines, offers, creatives) more efficiently by exploring less promising options less frequently. This means your popup campaigns reach optimal performance much quicker, translating directly into higher conversion rates. We leverage this for per-page headline optimization, ensuring your messaging is always fine-tuned to the specific content a user is viewing.

What We Learned from 10,000 Popup Impressions: Real-World Insights

Through analyzing data from tens of thousands of popup impressions across hundreds of sites, we've gathered crucial insights. One major takeaway is that timing is paramount: a poorly timed popup, no matter how good the offer, can be counterproductive. Nielsen Norman Group research consistently shows that intrusive popups harm user experience, but well-timed, relevant ones can be highly effective.

We also observed that the average conversion rate for well-implemented popups hovers around 3.09%, a figure consistent with Sumo's 2016 study. However, the top 10% of our popups achieve conversion rates of 9.28% or higher, largely due to the precision of the ExitSense ML model and dynamic content. On the 1,000+ sites running LeadYup popups, exit-intent on mobile typically needs a scroll-up + idle hybrid trigger because a simple 'mouse-out' event doesn't exist on touch devices. Pure time-on-page triggers are often too blunt, missing real intent signals.

What Modern AI/LLMs Add to how our exit-intent ML model actually works

The integration of advanced AI and Large Language Models (LLMs) significantly elevates our approach compared to legacy, rule-based popup tools. Here's how:

• Per-Page Copy Generation: Traditional tools require manual A/B testing of static copy. Our LLM dynamically generates popup copy tailored to the specific content on the page the user is viewing. This hyper-personalization, driven by understanding page context, leads to significantly higher engagement rates.
• Behavioral Signal Fusion via XGBoost: Instead of simple 'if/then' rules, we use sophisticated machine learning algorithms like XGBoost to fuse the 26 behavioral signals. This allows the model to identify complex, non-linear relationships between signals that indicate exit intent, leading to far more accurate and nuanced timing than any hand-coded rule set.
• Adaptive Learning with Thompson Sampling: As mentioned, Thompson sampling, powered by ML, continuously refines headline and offer effectiveness without the long wait times of traditional A/B tests. This means smaller businesses and indie founders can achieve optimization speeds previously only available to large enterprises with dedicated CRO teams. This is a core part of how our exit-intent ML model actually works.

The Trade-Offs: What Doesn't Always Work (and Why)

While AI-driven exit-intent is powerful, it's not a magic bullet. Overly aggressive popup frequency, even if perfectly timed, can still annoy users. Our model learns to detect 'popup fatigue' and can temporarily dial back triggers for repeat visitors. Furthermore, generic, untargeted offers will always underperform, regardless of timing. Even the best ML model can't fix a bad offer or an irrelevant message. The content of your popup still matters immensely. We've seen that popups offering generic '10% off' with no clear value proposition often perform worse than highly specific, problem-solving offers, even with optimal timing. Sometimes, a simple, clear call to action on a well-designed popup builder is more effective than an overly complex, personalized one if the core offer isn't compelling.

Roman Bootko

Founder & CEO, LeadYup

Roman has built lead-capture products since 2019, serving 1,000+ websites across 12 countries. He writes about exit-intent ML, popup conversion data, and the unsexy reality of growing SaaS from zero.